Background Subtraction Methods

Simple Gaussian

The background is modeled by a single multivariate Gaussian probability density function based on recent pixel values of the source image. The mean and covariance matrix of the Gaussian at each pixel is continuously updated using an on-line cumulative average. The pixels at each frame time are classified as foreground or background by calculating the Mahalanobis distance between the source and background model pixels, and comparing this distance to a threshold.

This model can detect moving objects with simple static backgrounds in controlled lighting situations. Due to the blind update method employed, the model is less sensitive to initial conditions but stationary foreground objects are eventually incorporated into the background.

See: Y. Benezeth, P.M. Jodoin,  B. Emile,  H. Laurent,  C. Rosenberger. Review and Evaluation of Commonly-Implemented Background Subtraction Algorithms. Proc. International Conference on Pattern Recognition, pp. 1-4, 2008.

Parameters

Sensitivity

Determines the sensitivity to changes in the background. Low values enhance the detection of objects in the scene, but also make the model more sensitive to noise.

Learning Rate

The rate at which the model adapts to changes in the video image. Low values correspond to a slowly adapting model. High values make the model adapt quickly to scene changes.

Noise Variance

Sets the minimum value of the variance for the Gaussian model. Higher values are recommended for videos with noisy images.

Fuzzy Gaussian

A modified version of the Gaussian model that uses a fuzzy classification rule and performs fuzzy on-line cumulative averages for the selective updating of the mean and the covariance matrix. The fuzzy selective updating of the background provides a better segmentation of stationary foreground objects compared to the simple Gaussian model.

See:  M. Sigari, N. Mozayani, H. Pourreza. Fuzzy Running Average and Fuzzy Background Subtraction: Concepts and Application, International Journal of Computer Science and Network Security, Vol. 8, No. 2, pp. 138-143, 2008.

Parameters

Sensitivity

Determines the sensitivity to changes in the background. Low values enhance the detection of objects in the scene, but also make the model more sensitive to noise.

Learning Rate

The rate at which the model adapts to changes in the video image. Low values correspond to a slowly adapting model. High values make the model adapt quickly to scene changes.

BG Threshold

Sets the threshold for the fuzzy classification rule. Low values tend to include more pixels in the detected foreground objects.

Noise Variance

Sets the minimum value of the variance for the Gaussian model. Higher values are recommended for videos with noisy images.

Mixture of Gaussians

Implements a classic multivariate Gaussian mixture model where every pixel is represented by a mixture of four Gaussian distributions. The modelling of the Gaussians is based on the Mahalanobis distance between the source and background model pixels. This model is designed to handle multimodal backgrounds with moving objects and illumination changes.

As in the simple Gaussian model, the blind update employed by the method makes it less sensitive to initial conditions but tends to integrate stationary foreground objects into the background.

See: T. Bouwmans, F. El Baf, B. Vachon. Background Modeling using Mixture of Gaussians for Foreground Detection – A Survey. Recent Patents on Computer Science 1, 3, pp. 219-237, 2008.

Parameters

Sensitivity

Determines the sensitivity to changes in the background. Low values enhance the detection of objects in the scene, but also make the model more sensitive to noise.

Learning Rate

The rate at which the model adapts to changes in the video image. Low values correspond to a slowly adapting model. High values make the model adapt quickly to scene changes.

BG Threshold

Sets the threshold that determines which of the distributions correspond to background pixels. High values are adequate for simple unimodal backgrounds. Lower values are recommended for complex backgrounds with moving objects.

Noise Variance

Sets the minimum value of the variance for the Gaussian models. Higher values are recommended for videos with noisy images.

Adaptive SOM

Also known as a Self-Organizing Background Subtraction Algorithm (SOBS), it adaptively models the background using a competitive neural network similar to the Kohonen Self-Organizing Map (SOM). For each pixel, a neuronal map consisting of 3x3 weight vectors is defined. The incoming source pixels are mapped to the weight vector that is closest according to a Euclidean distance metric, and the weight vectors in its neighbourhood are updated. The set of weight vectors act as a background model that is used for background subtraction in order to identify foreground pixels.

The model can handle scenes containing multimodal backgrounds with moving objects and gradual illumination changes. It employs a selective updating procedure that prevents the inclusion of stationary foreground objects into the background. In order to work properly a good initial background is recommended for training.

See: L. Maddalena, A. Petrosino. A Self-Organizing Approach to Background Subtraction for Visual Surveillance Applications. IEEE Transactions on Image Processing, Vol. 17, No. 7, pp. 1168-1177, 2008.

Parameters

Sensitivity

Determines the sensitivity to changes in the background. Low values enhance the detection of objects in the scene, but also make the model more sensitive to noise.

Learning Rate

The rate at which the model adapts to changes in the video image. Low values correspond to a slowly adapting model. High values make the model adapt quickly to scene changes.

Training Sensitivity

Determines the sensitivity to changes in the background during the training or calibration phase of the model, during which the neural network learns an initial background. High values are recommended for efficient learning.

Training Learning Rate

The rate at which the model adapts to changes in the video image during the training or calibration phase. High values are recommended for efficient learning.

Training Steps

Number of frames used for the training or calibration phase. More complex backgrounds usually require a greater number of training steps.

Fuzzy Adaptive SOM

A modified version of Adaptive SOM that uses a fuzzy rule to update the neural network background model. The fuzzy updating of the background helps to make the model more robust to illumination changes in the scene.

See: L. Maddalena, A. Petrosino. A fuzzy spatial coherence-based approach to background/foreground separation for moving object detection. Neural Computing & Applications, Vol. 19, No. 2, pp. 179-186, 2010.

Parameters

Sensitivity

Determines the sensitivity to changes in the background. Low values enhance the detection of objects in the scene, but also make the model more sensitive to noise.

Learning Rate

The rate at which the model adapts to changes in the video image. Low values correspond to a slowly adapting model. High values make the model adapt quickly to scene changes.

Training Sensitivity

Determines the sensitivity to changes in the background during the training or calibration phase of the model, where the neural network learns an initial background. High values are recommended for efficient learning.

Training Learning Rate

The rate at which the model adapts to changes in the video image during the training or calibration phase. High values are recommended for efficient learning.

Training Steps

Number of frames used for the training or calibration phase. More complex backgrounds usually require a greater number of training steps.

Laurence Bender